.

V

g

好

G

.

Given the very high quality of Professor Andrew Ng's lectures, I wanted to give this five stars, and I have given his previous classes five stars for that reason. You truly do learn a lot from what he teaches!

Sadly I have to downrate this particular class due to huge technical problems with submitting Jupyter notebook based assignments. In particular, week 4's face recognition assignment was marred by several issues. For one thing, the grader would often crash and report technical issues instead of grading the assignment. For another thing, the grader would also take 30 to 60 minutes to run, which is far longer than it takes to actually run the Jupyter notebook itself! Finally, and quite seriously, in order to get 30/30 on the face recognition assignment of week 4, I had to submit an INCORRECT answer due to a bug in the grader itself. At least for me these are primarily week 4 issues.

This class is still a terrific value and a valuable course for anyone wishing to study deep learning, and I am planning to make good use of these lessons. But I do believe that especially for the face recognition assignment it would have behooved the developers to test the assignment thoroughly before making it available to students and also to correct the substantial quality problems (i.e. failing a correct answer and accepting a wrong answer along with grader crashes) involving the grader. I hope these programming assignment quality problems do not appear again in the fifth course on recurrent networks because I am very much looking forward to that particular course given some of the projects I work on!

I was torn on whether to give this three stars only or four stars but decided that given the overall learning value I am still giving it four stars. But again I do hope the developers are reading these reviews and also the discussion forums regarding the major quality issues involving the face recognition programming assignment. The other assignments were very good and did not pose such ridiculous issues (my experience only.)

(09/2019)

Overall the courses in the specialization are great and provide great introduction to these topics, as well as practical experience. Many topics are explained clearly, with valuable field practitioners insight, and you are given quizzes and code-exercises that help deepen the understanding of how to implement the concepts in the videos. I would recommend to take them after the initial Andrew Ng ML course by Stanford, unless you have prior background in this topic.

There are a few shortbacks:

1 - the video editing is poor and sloppy. Its not too bad, but it’s sometimes can be a bit annoying.

2 - most of the exercises are too easy, and are almost copy-paste. I need to go over them and create variations of them in-order to strengthen my practical skills. Some exercises are quite challenging though (especially in course 4 and 5), and I need to go over them just to really nail them down, as things scale up quickly. Course 3 has no exercises as its more theoretical. Some exercises have bugs - so make sure to look at the discussion board for tips (the final exercise has a huge bug that was super annoying).

3 - there are no summary readings - you have to (re)watch the videos in order to check something, which is annoying. This is partially solved because the exercises themselves usually hold a lot of (textual) summary, with equations.

4 - the 3rd course was a bit less interesting in my opinion, but I did learn some stuff from it. So in the end it’s worth it.

5 - Slide graphics and Andrew handwriting could be improved.

6 - the online Coursera Jupyter notebook environment was a bit slow, and sometimes get stuck.

Again overall - highly recommended

Overall, the course is a great resource. The reasons why 4 stars and not 5:

Course material sometimes not as rigorous as one would expect: formulas for the same thing changing from one slide to the next (with the second slide not being correct), missing or erroneous indices like summation indices,... This is a bit unfortunate, as the course material is the one thing one would expect to be 100% accurate.

Quiz questions sometimes ambiguous when not outright confusing, making you wonder what the quiz author is really after. In some instances (luckily not too many), getting the right answer is as much about second-guessing what the quiz author intent is as it is about checking you have understood the course contents correctly. Sometimes you guess right, sometimes you don't! On average, there is one quiz question per quiz where this is the case, so, you end up submitting twice: the first time with the answer to the question that you think the quiz author meant, and the second time with the answer that the quiz author expects, given what he really meant.

Mentor support: this remark is not specific to this particular course, it is more of a general issue with Coursera courses. The Mentor system where mentors are supposed to help you out, while a great idea, just does not work as well as most learners can expect. There are situations where you can clearly see that mentors "pick" the forum questions they feel they can answer, but leave some others unanswered, either because the question does not interest them, or because they simply don't know.

Basically, last course in that specialization have not started yet. So, let me share feedback about current situation and my feelings about specialization here. Right now 4 courses released and they are all quite different if you will compare amount of practical exercises. This exercises are most important, but some of "weeks' doesn't have programming assignments at all. There is no homework here, and such weeks are quite boring. Python notebooks executed on very weak virtual machines and estimated execution time mentioned in notebooks seems VERY optimistic. You cannot debug tasks properly, and you couldn't play with models. And from other side you are not providing convenient way to just download task and run on my own server. Yes, you can say that a lot libraries required to run assignment's. But c'mon if you aimed on future machine learning engineers, please give them a chance to setup software that they will use every day. Another problem in very big amount of misprints in formulas in notebooks, which makes tasks more difficult in unnecessary places. Quiz questions also a very general, most of them doesn't require any thinking, just good memory.

Still, I think this is a one of the best specializations. But as for course complexity level you should be closer to J. Hinton level.

The ideas contained in this course are exceptional, and the delivery is also generally pretty good. I believe that because this is the beta delivery, there are a number of issues in the videos that are mostly minor annoyances (edits that lead to repeated portions of video, "first tries" being left in, and so on). Furthermore, I found the lab portions to be overly scaffolded. After this course and others, I have definitely learned many fundamentals of the topic, but there are steps in the implementation that I'm fully convinced I couldn't do on my own. Things like loading data, chopping up previously learned networks, and reformatting image data to be fed into a networks (to name a few) would stymie my efforts to train even a basic network. I understand that this was done in service of getting to the actual content, but the assignment that contained the IoU implementation and Yolo left me no more able to code Yolo than when I started. I feel that less of the code should be provided but copious hints offered. Alternatively, very little code given, but more leniency in the fora for students to share code ideas that they come up with.

Despite these misgivings, I eagerly await the RNN course!

Very good overview of convolutional neural networks. I especially liked the first weeks videos that explained the core concepts. It was interesting to then take these core concepts and show how they have been improved upon by adding batch norms, pooling layers and residual connections. Definitely the course to check out if you want to know the evolution of neural nets applied to image net. The YOLO algorithm is also covered for those interested in object recognition with bounding boxes, with the use case being self driving cars. Neural style transfer was interesting, though it's basically unusable on coursera due to lack of compute - get your own gpu for this I think.

Cons:

A lot of the videos had cutting issues so Andrew would appear to repeat himself - often with slightly different phrasing - annoying. I felt the assignments had less control as the coursera platform is not really designed for the workload required for neural networks. Lost connection to the jupyter notebooks a lot! Uses tensorflow which apparently isn't quite as good these days for research, though apparently is good in production environments?

If I could give 4.5 stars for this course, I'd do so.

'Convolutional Neural Networks' deserves a high rating. In my opinion the course is well designed, and taught by good presenter who has great experience. The course builds well, and the assignments show imagination. Prof Ng also did a good job of explaining and linking to recent developments.

At the same time, this course is correctly labelled as intermediate for difficulty. It took me far more than the advertised number of hours to complete each session.

So why not full marks? There were a number of mistakes and lack of clarity in some of the assignments that students complained about for years without correction. As well, a course that lists hundreds of thousands of students can certainly afford to have the prof re-record his lectures instead of printing corrections.

This course could also benefit by having mini-quizzes or even individual questions within lectures. There is often so much new material that it would be a good idea to divide the week into sections with a quiz after each.

Overall worthwhile, but there is room for improvement.

Andrew Ng is, as usual, a fantastic teacher. The quality of the content is stellar, like in the other courses in this series. However, after going through a few of these courses, I'm starting to resent the low production quality. The subtitles/transcripts are full of mistakes, to the point where they sometimes go for long stretches without a sentence that's transcribed correctly. Frustratingly, it feels like it would take just a quick pass by a human to fix their most glaring mistakes. (It's surprising how many ways a computer transcription system can misunderstand the term "ConvNets"). The end-of-week exercises are generally very good, but also plagued by annoying problems–insufficient tests that let mistakes slip through, a few exercises that assume knowledge that hasn't been presented adequately, and so forth.

All in all, this is another essential course from Andre Ng. However, it needs a few days of work to smooth out the rough edges.

Study materials good , have some errors but course is unique , programing assignment needs to be updated , great content , learned a lot about convolutions , stride , padding , max pooling , filters , channels , image recognition , face detection , NTS , and not to forget YOLO which (I would always thought how it happens)

Glad now I am able to see and elaborate underlying concept of all above and let others learn too

Sadly 2 downsides are :

1. Forum is mostly dead , I tried to ask 14 question and got only answered one , mentors not responding to question , one should always clarify even a single doubt , but not in case , so go for course if you have self pace and understanding , for me mentoring not worked , I had to google it !

2. Code needs to be updated for tensorflow 2.0 and up rather than tensorflow1 as company uses and requires tf 2.0 which have keras inbuilt implemented

So giving 4 stars

As usual, Andrew Ng presents concepts clearly and the homework assignments effectively reinforce those concepts.

Typically the biggest downside to these courses is the slightly clunky Coursera website and platform. However, in this case, the course content itself seems to have been slapped together in order to meet a promised deadline. The video edits are especially telling, with repeated segments (or, "do overs") where the repeated segment provides a correction to a botched segment, but then this botched segment is inexplicably left in the final video.

Still, Andrew Ng's courses are popular for a reason. His knowledge is unimpeachable, his ability to present concepts clearly, and importantly his friendly delivery makes him a charismatic figure. I've benefited from 5 of his courses to date and would likely enroll in any of his future offerings.

I wish those courses were available two years ago. A lot of useful information about convolutional networks, how they work and how they could be applied to real problems. I highly recommend this course for anyone interested in artificial intelligence and deep learning. The strong side of this course is that material is presented in clear and understandable form.

Course is new so you should expect some mistakes here and there. They will be ironed out with a time.

Practical tasks are really good. You get a possibility to use numpy, tensorflow and keras. However there is too much of emphasis on how to run tensorflow graphs (building graphs was not a problem, running them was), which requires a very specific knowledge and takes quite a lot of time searching for info.

Overall evaluation of this course is very good and I do recommend it.

While I though the last two courses of this specialization were a bit lackluster, this one was absolutely fantastic again. Definitely the best one of this specialization so far. The content was very in-depth and challenging in a very positive way.

Sadly, I'll have to subtract one star due to the grader of programming assignments not working well. Having to read through the discussion boards all the time to find workarounds for common bugs and grader issues, and then still have it fail because the grader is apparently down right now... didn't have too much fun with that. Some Jupyter Notebooks were already updated to a 3rd version since the few weeks of this courses release but would still not work correctly.

For the course on Sequence Models, please ensure a higher quality of programming assignments at release!

This course is really well thought and executed,all the recent algorithms that are being used in the industry are taught but this course in not for beginners i.e who are new to this domain without having any previous knowledge about tensors.I had a really hard time understanding the 'axis' part of most of the concepts as they are not really pointed out by the instructor and a vague idea is given about them in the assignment.There were few arts in the assignment where the cells were pre coded by the instructor and i couldnt understand whats happening in the cell so i just executed it and switched to the part where i had to code.It would have been better if little bit more information was given whats happening in that cell.Overall excellent course but few flaws like any other online course.

This course is very informative and helpful as was the other courses. What I found missing was after doing this course, i know what the individual modules/functions are doing but I am still pretty much confused in the full implementation of the different algorithms with CNNs. More focus can be given to students to let them code the whole system rather than just letting them implementing small parts. More infromative instructions could be added with the whole implementation but let the students code the whole system. It will retain in the student's memory for more time. and gives a broader view of the working implementation of the system. Finally, I am grateful to Prof. Andrew Ng and coursera to allow me to learn through this course with Financial Aid. Keep up the AWESOME work. :)

Of the courses in this specialisation, this was both the most interesting and the most challenging to date. There were positive and negative aspects to this. I found myself needing to look beyond the provided source material in order to fully understand what was going on, and whilst undertaking and submitting the assignments, spent almost as much time getting frustrated by issues with the grader as much as solving the problem at hand. The a number of the videos also seem to contain some editing issues (in terms of repeating the same small fragments twice, either that or the video player has an issue). However, I've been excited to learn about conv nets and their amazing applications and look forward to putting the learning into practise at some point in the near future.

Proud to have mastered this course. Great material, great teacher, challenging programming assignments and quizzes. Thank you!

Some optimisation ideas: There are quite many lectures without downloadable slides. In programming assignment 2 the links to www.tensorflow.org do not work. The Keras tutorial contains a misspelled loss function 'binary_cross_entropy'. Programming assignment 'Art Generation with Neural Style Transfer' does not provide a submit button. According to the discussion forums many of these weaknesses were reported three or so years ago but the weaknesses still exist today. This is a pity as it creates the impression that there is no understanding of the importance to perform good practise maintenance on the published material.

This course gave a very good high-level overview of the material, with lots of real-world examples. The programming exercises don't require a lot of Python knowledge, so long as you can use online search tools effectively to get command syntax and other tips. The starter code may seem a bit like it is leading you by the nose to the solution. But, given the time frame of the course, it is excellent for older learners who probably have less time to dedicate to the coursework (due to work and family constraints). The English transcripts are not very good, at times, from what I noticed of the top of them, as Andrew was speaking. Even though the subject matter is usually considered intimidating, it is presented in a very manageable way, here.

Good content, worth the money even considering the not so "fair" subscription policy and switching sessions, but as it's for profit business, it is understandable. W4 face recognition assignment is very buggy, also the whole website seems to be really slow in general. What's even worse is that many people complained about it for months yet neither Coursera nor teaching stuff from Stanford did anything to fix it. Making unfinished product and not fixing the bugs in such long period of time doesn't reflect well on Prof. Andrew Ng, since after all he is co-founder of the platform. Regardless of that I'm still very grateful to him and the stuff from Stanford for the high quality of the material presented here and for so many great exercises.

The course is wonderful and explains almost every thing very eloquently.However, the course should mention Tensorflow and Keras as prerequisites for the course since they are heavily used in the assignments. Although there has been an effort to explain how tensorflow works, not enough details and not enough time was provided to get a good grasp of tensorflow and keras. So a beginner in tensorflow and keras will find it hard to stomach the essence of how to code these systems even if she completes the assignments with ease.A small course on just tensorflow and keras may be introduced in the specialization before this course. That would greatly increase the efficacy of this course and will also make the course more interesting.